CN113962472B - Short-time prediction method for space-time double-attention subway passenger flow based on GAT-Seq2Seq model - Google Patents
Short-time prediction method for space-time double-attention subway passenger flow based on GAT-Seq2Seq model Download PDFInfo
- Publication number
- CN113962472B CN113962472B CN202111278744.6A CN202111278744A CN113962472B CN 113962472 B CN113962472 B CN 113962472B CN 202111278744 A CN202111278744 A CN 202111278744A CN 113962472 B CN113962472 B CN 113962472B
- Authority
- CN
- China
- Prior art keywords
- time
- subway
- passenger flow
- data
- gat
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 46
- 239000013598 vector Substances 0.000 claims abstract description 58
- 239000011159 matrix material Substances 0.000 claims abstract description 33
- 238000012549 training Methods 0.000 claims abstract description 20
- 230000007246 mechanism Effects 0.000 claims abstract description 17
- 238000013528 artificial neural network Methods 0.000 claims abstract description 15
- 238000012360 testing method Methods 0.000 claims abstract description 14
- 125000004122 cyclic group Chemical group 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 6
- 230000006870 function Effects 0.000 claims description 28
- 238000004364 calculation method Methods 0.000 claims description 19
- 230000008569 process Effects 0.000 claims description 12
- 238000012795 verification Methods 0.000 claims description 11
- 230000008859 change Effects 0.000 claims description 9
- 230000002159 abnormal effect Effects 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000013527 convolutional neural network Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000011156 evaluation Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 230000001131 transforming effect Effects 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 238000006467 substitution reaction Methods 0.000 claims description 2
- 238000010586 diagram Methods 0.000 description 7
- 238000000605 extraction Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 230000009466 transformation Effects 0.000 description 3
- 238000011161 development Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000007726 management method Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Human Resources & Organizations (AREA)
- Software Systems (AREA)
- Strategic Management (AREA)
- Tourism & Hospitality (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Economics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Evolutionary Biology (AREA)
- Development Economics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Entrepreneurship & Innovation (AREA)
- Primary Health Care (AREA)
- Educational Administration (AREA)
- Game Theory and Decision Science (AREA)
- Quality & Reliability (AREA)
- Operations Research (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a short-time prediction method of space-time double-attention subway passenger flow based on a GAT-Seq2Seq model, which comprises the following steps: acquiring subway passenger flow data; data preprocessing and data set dividing; generating a graph structure network according to the geographical position of the site and the operation route; constructing a GAT model based on multi-head attention, inputting feature vectors in a training set, and calculating the spatial relevance of sites by combining a site map structure; transmitting the feature vector output by the GAT into a Seq2Seq model based on time sequence attention, extracting the time correlation of the passenger flow, calculating the mean square error by using a training set, and adjusting the edge weight matrix of the graph structure in the GAT and the cyclic neural network parameters in the Seq2Seq model; the test set is used to predict and evaluate the model. The invention solves the problem that the accuracy of the prediction result is low because the existing prediction model only searches the characteristics from the short time dimension by utilizing a space-time dual-attention mechanism.
Description
Technical Field
The invention relates to a short-time prediction method for space-time double-attention subway passenger flow based on a GAT-seq2seq (Graph Attention Network & Sequence to sequence) model, and belongs to the technical field of subway passenger flow prediction.
Background
With the continuous development of social economy, the level of urbanization is continuously improved, and the subway guarantees the basic demands of residents on life and travel. However, due to the influence of the peaks in the morning and evening and weather, subway passenger flows show periodic fluctuation, so that in-station congestion conditions in certain time periods are caused, management difficulty of staff is increased, subway departments can be assisted to reasonably distribute and schedule subway transportation capacity in the whole city range through accurate prediction of the subway passenger flows, early warning is achieved for large-scale crowd intensive activities, planning and layout are conducted on future subway lines and stations, and huge help can be provided for management of whole subway networks and even whole city traffic.
The influence factors of subway passenger flow are various and wide in range, and comprise weather factors (such as rainfall, snowfall, temperature and the like), different daily time periods, holidays, space geographic positions of sites and the like. The historical data acquisition of the traditional subway traffic passenger flow prediction can be obtained only by relying on a large amount of manpower and material resources to conduct long-time investigation in the field, and the final prediction result can only be used for observing daily subway flow change and cannot cope with emergency situations caused by large-scale crowd activities. Moreover, the prediction method cannot take the correlation strength among all influence factors into consideration after being quantized, and extracts daily trends, monthly trends, quarterly trends and annual trends from the time angle to roughly predict the passenger flow, so that the prediction method has no universality.
Along with the continuous development of data mining and deep learning, the technologies of convolution calculation, feature extraction, error back propagation and the like are mature gradually in the field of passenger flow prediction, and the change rule of a neural network model for a long period of time in historical data can be established by means of the efficient computing capability of a computer, so that the prediction accuracy is improved, but the spatial relation of passenger flow on the distribution of each site is still not fully considered, or the influence weight value of each adjacent site to the predicted site is not reasonably distributed. Meanwhile, the LSTM model-based prediction method omits a large number of time sequence features by adopting simple forgetting, adding, updating and other operation units when extracting the time sequence rule, so that the accuracy of the whole prediction model is lower.
Disclosure of Invention
The technical problems to be solved by the invention are as follows: the time-space double-attention subway passenger flow short-time prediction method based on the GAT-Seq2Seq model is provided, attention mechanisms are considered from the space and time, the problem that attention mechanisms exist when the existing main stream prediction model fails to consider time-space angles, and only time sequence rule extraction in a short time is considered, so that the prediction result precision is low is solved. Meanwhile, the existing time sequence attention mechanism is improved in the Seq2Seq model part, so that the model can process passenger flow prediction in a long period range in parallel.
The invention adopts the following technical scheme for solving the technical problems:
A short-time prediction method of space-time double-attention subway passenger flow based on a GAT-Seq2Seq model comprises the following steps:
step 1, collecting subway passenger flow related data, which comprises the following steps: subway station basic information, subway station historical passenger flow data, historical meteorological data and holiday data;
Step 2, preprocessing collected subway passenger flow related data, sequentially ordering the history passenger flow data and the history meteorological data according to time to generate time sequence data, dividing a day into a plurality of time periods, counting subway incoming passenger flow, rainfall, snowfall and temperature data in different time periods, normalizing the counted data to obtain a data set, and dividing the data set into a training set, a verification set and a test set according to the proportion of 3:1:1;
step 3, taking all subway stations as vertexes of the graph structure, judging whether edges exist between any two vertexes through a subway running line and a set distance threshold value, generating a subway station graph structure G, initializing edge weights between the vertexes by adopting an Xavier method, and obeying uniform distribution;
step 4, constructing a graph attention convolutional neural network model GAT, and constructing a feature matrix by using the training set in the step 2 The feature matrix/>As input of the network model GAT, calculating the attention coefficient between each vertex in the graph structure G in step 3, and then weighting and summing to obtain a feature matrix/>, which considers the influence of the neighborhood site
Step 5, constructing a time sequence to time sequence model Seq2Seq based on time attention, and obtaining the feature matrix in the step 4Inputting the time sequence to a time sequence model Seq2Seq, after processing by an encoding layer and a decoding layer, calculating the root mean square error between the predicted value and the actual value of the subway incoming passenger flow, adjusting the super-parameters of the circulating neural network unit in the time sequence to the time sequence model Seq2Seq according to the root mean square error and a training set, and further optimizing the super-parameters by utilizing a verification set, thereby obtaining a trained GAT-Seq2Seq model;
And 6, inputting a test set into the trained GAT-Seq2Seq model, predicting the passenger flow of subway stations in each future period, and evaluating the trained GAT-Seq2Seq model according to a prediction result.
As a preferred embodiment of the present invention, in the step 1, the subway station basic information includes: longitude and latitude geographic coordinates of subway stations and subway operation line data; the subway station history passenger flow data comprises: station names or station serial numbers, card swiping inbound time and inbound passenger flow; the historical meteorological data includes: whether to rainfall, whether to fall snow, the rainfall, the snowfall and the temperature; holiday data includes: whether the day of the week and whether the holiday, wherein whether the day of the week is converted into seven parameters using a thermal independent code, namely: whether Monday, tuesday, friday, saturday, and Sunday.
As a preferable scheme of the invention, the specific process of the step 2 is as follows:
step 21, deleting the data with abnormal or missing subway station names in the related data of the subway passenger flow by adopting an integral removing method, and processing the data with the missing or abnormal subway station names in the subway card swiping and station entering time, station entering passenger flow, rainfall, snowfall and temperature by adopting a forward filling substitution method;
Step 22, sorting historical passenger flow data of subway stations according to time sequence, and setting the time period length l to be 1 hour, dividing one day into 24 time periods, and respectively counting the passenger flow, rainfall and snowfall of the subway stations in each time period;
Step 23, calculating the average temperature in each period according to the acquired temperature data, wherein the maximum temperature and the minimum temperature in each period respectively take the maximum value and the minimum value of all the temperatures in the period;
Step 24, normalizing the subway card-swiping and arrival passenger flow, the rainfall, the snowfall, the average temperature, the highest temperature and the lowest temperature to obtain a data set;
Step 25, the data set is divided into a training set, a verification set and a test set according to the ratio of 3:1:1.
As a preferable scheme of the invention, the specific process of the step3 is as follows:
Step 31, taking all subway stations as vertexes of a graph structure, knowing longitude and latitude coordinates of each subway station, and for any vertex A, searching a circular range with the vertex A as a circle center and the radius of 2.5Km, wherein if a certain subway station simultaneously meets the following two conditions: (1) The subway station is in a circular range, (2) a subway operation reachable line exists between the subway station and the vertex A or the actual walking time between the subway station and the vertex A is less than 10 minutes; then consider that there is an edge between the subway station and vertex a, otherwise there is no edge;
step 32, repeating step 31 for all vertices to obtain the subway station map structure G.
As a preferred scheme of the present invention, the specific process of the step 4 is as follows:
Step 41, the input of the graph-meaning force convolutional neural network model GAT is a four-dimensional feature matrix Wherein D is the total time span of training set data, T is the number of time periods in the day, N is the number of subway stations, F is the feature vector dimension, and then the feature matrix/>The method comprises the following steps:
wherein, At the t-th period of the D-th day, F feature vectors affecting passenger flow change of subway stations numbered n, D e [1, D ], t e [1, t ], n e [1, n ], d=1096, t=24, n=83, f=16;
Step 42, for vertex s i of the graph structure, calculating the correlation coefficient between vertex s i and neighbor vertex s j For evaluating the correlation between s i and s j,/>The calculation formula is as follows:
Where s j denotes the neighbor vertices that have edges with s i, Representing a set of neighbor vertices with edges s i, W representing an edge weight matrix, |operation representing a lateral concatenation operation, a (·) representing a transformation function mapping high-dimensional features to a real number,/>F feature vectors which affect passenger flow change of subway stations with numbers of i and j at the t time period of the d day are respectively represented;
step 43, correlation number Normalization yields the attention coefficient/>The specific calculation formula is as follows:
wherein LeakyReLU (·) is a normalized activation function, s k represents neighbor vertices with s i existing edges;
step 44, according to the attention coefficient Feature vector/>Weighted summation yields a new eigenvector/>The specific calculation formula is as follows:
Wherein σ (·) is the calculation function;
Step 45, generating different edge weight matrix W and a (& gt) functions by considering M groups of different edge weight matrix W and a (& gt) functions And then will beAnd transversely splicing, and finally obtaining an average value, wherein the specific calculation formula is as follows:
wherein, Representing the attention coefficients calculated using the m-th set of weights W and a (·) functions, W m representing the m-th set of edge weight value matrices;
Step 46, feature vector for each period within the total span of time D Steps 42-45 are performed to obtain a new feature vector/>New feature vector/>According to step 41/>Is spliced to generate GAT output matrix/>
As a preferable scheme of the invention, the specific process of the step 5 is as follows:
step 51, constructing an encoder, wherein the encoder adopts a cyclic neural network unit to obtain the feature matrix obtained in the step 4 The data of each day of (a) is tiled and unfolded, and is marked as X, X= { X 1,x2,x3,…,xT }, D X vectors are taken, a cyclic neural network unit is marked as p, a hidden layer state variable h t at the time t is only related to a hidden layer state variable h t-1 at the time t-1 and an input X t at the time t, and the formula is expressed as follows:
ht=p(xt,ht-1)
constructing D parallel encoders, i.e. constructing an encoder comprising T time periods for each of D days, for generating a unified context vector c, then the context vector c is populated with hidden layer state variables by the encoder And (3) transforming to obtain:
wherein q (·) is the transform function;
Step 52, a decoder is constructed by considering the attention mechanism, and the decoder adopts a cyclic neural network unit and is marked as g; the hidden layer state variable s t′ at time t 'of the decoder is related only to the hidden layer state variable s t′-1 at time t' -1, the context vector c t′ at time t '-1, the output variable y t′-1 at time t' -1, the formula is as follows:
st′=g(yt′-1,ct′,st′-1)
The context vector c t′ of the decoder at the time t' is calculated by dividing the context vector c t′ into two parts, wherein the first half part only considers the time sequence information of 24 time periods of the last day, and the second half part considers the information of the same prediction time period under all previous days, and the specific calculation formula is as follows:
wherein, Hidden layer state variable representing time t of day D in encoder,/>Hidden layer state variable representing the time t' of the d-th day in the encoder,/>Time attention coefficient representing the state at time t' in the decoder and the state at time t on day D in the encoder,/>Time attention coefficient representing the state at time t 'in the decoder and the state at time t' on day d in the encoder, time attention coefficient/>The calculation formula is as follows:
wherein, A correlation coefficient representing a state at time t' of the decoder and a state at time t of the D-th day of the encoder; /(I)A correlation coefficient representing a state at time t 'of the decoder and a state at time t' of the d-th day of the encoder;
all context vectors c t′ corresponding to the time t' in the decoder are calculated by the above formula and are used for calculating the predicted data y t′:
yt′=σ(yt′-1,ct′,st′);
And step 53, using the root mean square error RMSE as a loss function of the model, using an Adam SGD optimizer to adjust related parameters in the GAT-Seq2Seq model according to the training set, and performing super-parameter optimization through the verification set, thereby obtaining the trained GAT-Seq2Seq model.
As a preferred scheme of the present invention, the specific process of the step6 is as follows:
step 61, migrating the trained GAT-Seq2Seq model to a test set, determining a prediction step length l to be 1 hour, and predicting subway station passenger flow in a future period by using the trained GAT-Seq2Seq model;
Step 62, calculating a root mean square error RMSE of the predicted value and the actual value for model performance evaluation, wherein the root mean square error RMSE is as follows:
wherein R is the total number of predicted samples, y r is the actual passenger flow volume, To predict passenger flow.
Compared with the prior art, the technical scheme provided by the invention has the following technical effects:
1. the method considers the mutual influence of subway stations on passenger flows in space and geographic positions based on the graph convolution neural network, and realizes influence degree weight distribution on adjacent stations by combining an attention mechanism.
2. The invention adds a time attention mechanism on the basis of adopting a time sequence to time sequence model, innovatively adopts a multi-encoder fusion mode, and generates context vectors for different hidden state units of a decoder by combining historical feature vectors of each day in a data set under the time period when predicting the passenger flow of the time period.
3. The invention combines space-time characteristics, adopts attention mechanisms in space and time, particularly improves a time sequence attention mechanism in a Seq2Seq model, so that the model can process passenger flow prediction in an ultra-long period range in parallel, and the improvement and application obviously improve the precision of a prediction model and predict daily regular passenger flow change of subway stations accurately; for large-scale travel emergencies, the model can also be combined with passenger flow data of adjacent sites to make short-time accurate predictions.
Drawings
FIG. 1 is a logic diagram of the overall flow of the method of the present invention;
FIG. 2 is a diagram of a detailed model operation structure of the method of the present invention;
FIG. 3 is a schematic diagram of a data preprocessing flow for the method of the present invention;
FIG. 4 is a block diagram of a Seq2Seq model encoder of the method of the present invention;
FIG. 5 is a graph comparing the predicted results of the model of the present invention with actual data.
Detailed Description
Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings. The embodiments described below by referring to the drawings are exemplary only for explaining the present invention and are not to be construed as limiting the present invention.
The method and the device have the core ideas of extracting the spatial correlation and the time correlation of the internal characteristics of the data, and predicting the passenger flow of the future period after combining the spatial correlation and the time correlation. In the extraction of the spatial correlation characteristics, the interaction between adjacent stations is determined by using a multi-head attention mechanism through the geographical position and the route of the stations by using the current advanced graph convolution neural network. In time-related feature extraction, the improved time sequence-to-time sequence model is used for extracting influence features at the time level, namely, passenger flow data of the first time period can influence passenger flow prediction results of the later time period.
The data used in the embodiment of the invention are gate passenger flow record data of all stations of a subway system 2018, 6 months and 2021, 6 months of a city abroad, and weather and air quality data of all areas of the city disclosed by the city weather department. Subway station position data and operation line data are derived from the public of the urban transportation departments. And finally, homemade holiday information of 2018-2021 comprises the information of the national fixed holidays, working days and non-working days.
Referring to fig. 1, which is a general flow chart of the space-time dual-attention subway short-time passenger flow prediction method based on the GAT-Seq2Seq model, fig. 2 is a detailed operation structure diagram of each part of the present invention, and the method of the present invention is further described below with reference to fig. 1 and 2, and includes the following steps:
Step S1: collecting subway passenger flow related data, wherein the data comprise: subway station basic information, station historical passenger flow data, historical meteorological data and holiday data. Wherein the site basic information includes: the longitude and latitude geographic coordinates of each subway station and subway operation line data. The historical passenger flow card swiping data comprises: station name (or station number), gate card swiping time and number of people entering the station at the time. The meteorological data includes: whether to rainfall, whether to fall snow, the amount of rainfall, the amount of snowfall, and the temperature. Holiday data includes: whether day of week, whether legal holidays. Wherein, whether the day of the week parameter adopts the hot independent coding is converted into seven parameters: whether Monday, tuesday, friday, saturday, and Sunday.
Step S2: and (5) preprocessing data. And (3) generating time-series data by sorting subway card swiping data and meteorological data according to time, and counting the total subway passenger flow, the rainfall, the snowfall, the average temperature, the highest temperature and the lowest temperature in different time periods, wherein the time period is every 1 hour, and the data are normalized and an input characteristic matrix H is generated. The data set is divided into a training set, a verification set and a test set according to the ratio of 3:1:1.
The specific flow of data preprocessing and partitioning is further explained in conjunction with fig. 3:
step S21: and deleting the whole row of related data by adopting a whole deletion method for the data with abnormal or missing station names in the historical passenger flow data. And the front filling replacement method is adopted for the missing value and the abnormal value of the card swiping time, the number of people in the station, the rainfall, the snowfall and the temperature data.
Step S22: the historical passenger flow card swiping data is arranged according to the positive sequence of the time sequence, the length l (unit: hour) of the predicted time period is set to be 1 hour, the number T of the time period is calculated,And counting the number of people who enter the station, the rainfall and the snowfall in 24 time spans respectively.
Step S23: the average value in the period is calculated for the temperature data in the meteorological data at the same time span t=24 in step S22. The maximum temperature data is selected as the maximum value in the period, and the minimum temperature is selected as the minimum value in the period. The data after pretreatment are shown in table 1.
TABLE 1 data structure after pretreatment
Step S24: normalization processing is carried out on the pretreated data of the number of people entering the station, the rainfall, the snowfall, the average temperature, the highest temperature and the lowest temperature, the data are mapped between [0,1], and the normalization operation formula is as follows:
step S25: the normalized data set is divided into a training set, a verification set and a test set according to the proportion of 3:1:1.
Step S3: generating a subway station diagram structure G, wherein each station is a vertex of the diagram structure, and determining whether edges exist between any two stations through a subway running line and a distance threshold value. The graph structure is an infrastructure for learning the spatial influence between adjacent sites, and the edge weights W between the vertexes are initialized by adopting an Xavier method and are subjected to uniform distribution.
The specific flow of generating the site map structure is as follows:
Step S31: all subway stations are vertexes of a graph structure, the longitude and latitude coordinates of each station are known, any one vertex is selected, the radius is set to be 2.5Km, a circular range is found, and if the rest stations meet the following two conditions, (1) the stations are in the circular range, (2) the actual walking time between the stations and the subway running reachable line of the vertex or the subway running reachable line is less than 10 minutes, and the stations meeting the conditions are considered to have edges with the vertex.
Step S32: repeating the steps for all the vertexes to obtain the complete graph structure of the subway station.
Step S4: constructing a graph attention convolution neural network model for learning the geospatial relevance between subway stations, and inputting a feature matrix by using the graph structure G in the step S3 and the training set in the step S2Calculating the attention coefficient alpha among the vertexes in the step S31, and then weighting and summing to obtain a feature matrix/>, wherein the feature matrix/>, after influence of the neighborhood site is considered
The specific operation flow of the chart attention convolution layer comprises the following steps:
step S41: the input to the graph attention convolution network is a four-dimensional feature matrix Where D is the time span of the training set data (i.e., total days d=1096 days), T is the number of daily periods defined in step S22 (t=24), N is the number of subway stations (n=83), and F is the feature vector dimension (number of historical stops, whether to rain, whether to snow, rainfall, snowfall, average temperature, maximum temperature, minimum temperature, day of week parameters after heat-independent encoding, whether to legal holidays, 16 parameters in total, i.e., f=16). The input feature matrix is:
wherein, F (f=16) feature vectors of subway station numbered n, expressed as the nth period on day d, affecting passenger flow change, to/>For example,/> The 16 feature vectors affecting passenger flow change for subway station number 1 are expressed as the first time period on day 1 (12 a.m. to 1 a.m.).
Step S42: calculating a correlation coefficientFor the vertex s i, the correlation coefficient between the vertex and the neighbor point is calculated one by one, and is used for evaluating the correlation between s i and s j, and the calculation formula is as follows:
Wherein the method comprises the steps of S j is represented as the neighbor vertex with the edge of S i, resulting from the graph structure in step S31. In the formula, a shared parameter W is used for carrying out dimension increment on the vertex feature vector in a linear mapping mode, the operation is that the feature vectors of s i and s j after modification are subjected to transverse splicing operation, and finally, high-dimensional features are mapped to a real number through a (·) function operation.
Step S43: normalization of the correlation coefficient to obtain the attention coefficientThe specific calculation formula is as follows:
The normalized activation function is LeakyReLU (·) and takes the characteristic of the function into consideration that the characteristic of the function can keep the characteristic of the node s i itself, but not be influenced by the neighbor nodes, and the specific complete formula is as follows:
Step S44: according to the calculated attention coefficient, the feature vectors are weighted and summed to output new feature vectors (fused with neighborhood feature information), and the formula is as follows:
Step S45: the accuracy is enhanced by adopting a multi-head attention mechanism, namely, M groups of different weights W and a (& gt) functions are considered to generate different functions Will/>And transversely splicing, and finally obtaining an average value, wherein the specific formula is as follows:
Step S46: feature vector for each period (total period t=24) within the total span of data time (total days d=1096) The new feature vector/> is output through the stepsAccording to step S41/>Is spliced to generate GAT layer output matrix/>
Step S5: time sequence-to-time sequence models based on time attention are constructed, and data correlation at a time level is learned. Using the feature matrix obtained in step S4And inputting the parameter into the Seq2Seq model, and calculating the mean square error (RMSE) between the passenger flow predicted value and the actual value after the passenger flow predicted value and the actual value are processed by the coding layer and the decoding layer, so as to adjust the parameter of the cyclic neural network unit in the Seq2Seq model.
The specific flow of the time sequence to time sequence model comprises the following steps:
Step S51: an encoder module is constructed, the encoder adopts a cyclic neural network, the encoder is used for sequence coding of information, and sequence information with any length can be coded into a context vector c. The encoder model will be described with reference to fig. 4, in which the input vector is the feature matrix in step S46 Is spread out and denoted as X, x= { X 1,x2,x3,…,xT }, and there are a total of d=1096 (total days) X vectors. The hidden layer variable h t at the time of p and t is marked as p by the cyclic neural network unit, and the hidden layer state variable h t-1 at the time of t-1 is only related to the input x t at the time of t, and can be expressed by the following formula:
ht=p(xt,ht-1)
In the current prior art, the context vector c of the encoder output is typically transformed (as shown in the following formula) by all hidden layer variables h t of the encoder. Where q (·) is a transformation function, such as a summing function or an averaging function, etc.
c=q(h1,h2,…,hT)
If all the input vectors are unfolded and then transmitted into the model, D.T hidden layer states are generated, so that the calculated amount of the model is greatly increased, and therefore, the invention provides a parallel operation mode to improve the efficiency. Since there are a total of d=1096X input vectors, then D parallel encoders can be constructed, i.e. an encoder comprising 24 time periods for each of 1096 days, for generating a unified context vector c, the lower equation meaning that the context vector c can be populated with hidden state variables by the encoderAnd (3) transforming to obtain:
Step S52: the decoder module is constructed taking into account the attention mechanism, the purpose of which is to allow the decoder to use different context vectors c t′ at different times, since the attention mechanism is introduced, the context vector c in step S51 needs to be recalculated and decomposed into different c t′ in new theory. The decoder also adopts a cyclic neural network, which is marked as g; the hidden layer state variable s t′ at time t 'of the decoder is related only to the hidden layer state variable s t′-1 at time t' -1, the context vector c t′ at time t '-1, the output variable y t′-1 at time t' -1, the formula is as follows:
st′=g(yt′-1,ct′,st′-1)
In the current attention mechanism, the context vector c t′ of the decoder at time t' can be generally represented by a weighted sum of t hidden layer state variables h t in the encoder by the time attention coefficient α t′t, with the following formula:
However, this type of attention mechanism has an excessively long time span in this example, and repeatedly retains irrelevant timing information (the characteristic information of 24 periods of 1096 days is retained), so the context vector c t′ is creatively calculated by dividing the context vector c into two parts (the formula is as follows), the first half is the timing information of 24 periods of the last day only is considered, and the second half is the information of the same prediction period under all previous dates. For example, when predicting traffic at 8 to 9 of 21 year 6 month 2 days, the context vector c t′ needs to consider state variables of 24 periods of 21 year 6 month 2 days and state variables of all 8 to 9 of 18 year 6 month 1 day to 21 year 6 month 1 day, which can reduce the influence of useless data on the context vector while maintaining key timing information.
Wherein the time attention coefficientThe calculation formula is as follows, and it should be noted that the attention mechanism considers that all states h t of the D encoders are filtered:
/>
all context vectors c t′ corresponding to time t' in the decoder can be calculated by the above formula, and used for calculating the predicted data y t′, wherein σ (·) is a calculation function:
yt′=σ(yt′-1,ct′,st′)
Step S53: the root mean square error RMSE is used for obtaining a loss function principle serving as a model, an Adam SGD optimizer is used for adjusting related parameters (a weight matrix W in step S3 and a transformation function a T in step S43) in a prediction model according to training set data, and super-parameter optimization is carried out through verification set data in the later period of training, so that generalization capability of the whole model is improved, and a more accurate result can be obtained in unknown data by a final model.
Step S6: and inputting the test set data into a trained model, predicting the passenger flow of subway stations in each period, and carrying out model evaluation according to the prediction result.
The specific flow of testing and evaluating the model includes:
Step S61: and (3) migrating the trained GAT-Seq2Seq model in the step (S4) and the step (S5) to test set data, determining a proper prediction step length l to be 1 hour, and predicting the station passenger flow of the future period by using the model and the data.
Step S62: and calculating a Root Mean Square Error (RMSE) of the predicted value and the actual value for evaluating the performance accuracy of the model, wherein the RMSE is expressed as follows:
wherein R is the total number of predicted samples, y r is the actual passenger flow volume, To predict passenger flow, the final prediction result pair is shown in fig. 5.
The above embodiments are only for illustrating the technical idea of the present invention, and the protection scope of the present invention is not limited thereto, and any modification made on the basis of the technical scheme according to the technical idea of the present invention falls within the protection scope of the present invention.
Claims (5)
1. A short-time prediction method of space-time double-attention subway passenger flow based on a GAT-Seq2Seq model is characterized by comprising the following steps:
step 1, collecting subway passenger flow related data, which comprises the following steps: subway station basic information, subway station historical passenger flow data, historical meteorological data and holiday data;
Step 2, preprocessing collected subway passenger flow related data, sequentially ordering the history passenger flow data and the history meteorological data according to time to generate time sequence data, dividing a day into a plurality of time periods, counting subway incoming passenger flow, rainfall, snowfall and temperature data in different time periods, normalizing the counted data to obtain a data set, and dividing the data set into a training set, a verification set and a test set according to the proportion of 3:1:1;
step 3, taking all subway stations as vertexes of the graph structure, judging whether edges exist between any two vertexes through a subway running line and a set distance threshold value, generating a subway station graph structure G, initializing edge weights between the vertexes by adopting an Xavier method, and obeying uniform distribution;
step 4, constructing a graph attention convolutional neural network model GAT, and constructing a feature matrix by using the training set in the step 2 The feature matrix/>As input of the network model GAT, calculating the attention coefficient between each vertex in the graph structure G in step 3, and then weighting and summing to obtain a feature matrix/>, which considers the influence of the neighborhood siteThe specific process is as follows:
Step 41, the input of the graph-meaning force convolutional neural network model GAT is a four-dimensional feature matrix Wherein D is the total time span of training set data, T is the number of time periods in the day, N is the number of subway stations, F is the feature vector dimension, and then the feature matrix/>The method comprises the following steps:
wherein, At the t-th period of the D-th day, F feature vectors affecting passenger flow change of subway stations numbered n, D e [1, D ], t e [1, t ], n e [1, n ], d=1096, t=24, n=83, f=16;
Step 42, for vertex s i of the graph structure, calculating the correlation coefficient between vertex s i and neighbor vertex s j For evaluating the correlation between s i and s j,/>The calculation formula is as follows:
Where s j denotes the neighbor vertices that have edges with s i, Representing a set of neighbor vertices with edges s i, W representing an edge weight matrix, i operation representing a lateral stitching operation, a (-) representing a transform function mapping high-dimensional features to a real number,F feature vectors which affect passenger flow change of subway stations with numbers of i and j at the t time period of the d day are respectively represented;
step 43, correlation number Normalization yields the attention coefficient/>The specific calculation formula is as follows:
wherein LeakyReLU (·) is a normalized activation function, s k represents neighbor vertices with s i existing edges;
step 44, according to the attention coefficient Feature vector/>Weighted summation yields a new eigenvector/>The specific calculation formula is as follows:
Wherein σ (·) is the calculation function;
Step 45, generating different edge weight matrix W and a (& gt) functions by considering M groups of different edge weight matrix W and a (& gt) functions Will/>And transversely splicing, and finally obtaining an average value, wherein the specific calculation formula is as follows:
wherein, Representing the attention coefficients calculated using the m-th set of weights W and a (·) functions, W m representing the m-th set of edge weight value matrices;
Step 46, feature vector for each period within the total span of time D Steps 42-45 are performed to obtain a new feature vector/>New feature vector/>According to step 41/>Is spliced to generate GAT output matrix/>
Step 5, constructing a time sequence to time sequence model Seq2Seq based on time attention, and obtaining the feature matrix in the step 4Inputting the time sequence to a time sequence model Seq2Seq, after processing by an encoding layer and a decoding layer, calculating the root mean square error between the predicted value and the actual value of the subway incoming passenger flow, adjusting the super-parameters of the circulating neural network unit in the time sequence to the time sequence model Seq2Seq according to the root mean square error and a training set, and further optimizing the super-parameters by utilizing a verification set, thereby obtaining a trained GAT-Seq2Seq model; the specific process is as follows:
step 51, constructing an encoder, wherein the encoder adopts a cyclic neural network unit to obtain the feature matrix obtained in the step 4 The data of each day of (a) is tiled and unfolded, and is marked as X, X= { X 1,x2,x3,…,xT }, D X vectors are taken, a cyclic neural network unit is marked as p, a hidden layer state variable h t at the time t is only related to a hidden layer state variable h t-1 at the time t-1 and an input X t at the time t, and the formula is expressed as follows:
ht=p(xt,ht-1)
constructing D parallel encoders, i.e. constructing an encoder comprising T time periods for each of D days, for generating a unified context vector c, then the context vector c is populated with hidden layer state variables by the encoder And (3) transforming to obtain:
wherein q (·) is the transform function;
Step 52, a decoder is constructed by considering the attention mechanism, and the decoder adopts a cyclic neural network unit and is marked as g; the hidden layer state variable s t′ at time t 'of the decoder is related only to the hidden layer state variable s t′-1 at time t' -1, the context vector c t′ at time t '-1, the output variable y t′-1 at time t' -1, the formula is as follows:
st′=g(yt′-1,ct′,st′-1)
The context vector c t′ of the decoder at the time t' is calculated by dividing the context vector c t′ into two parts, wherein the first half part only considers the time sequence information of 24 time periods of the last day, and the second half part considers the information of the same prediction time period under all previous days, and the specific calculation formula is as follows:
wherein, Hidden layer state variable representing time t of day D in encoder,/>Hidden layer state variable representing the time t' of the d-th day in the encoder,/>Time attention coefficient representing the state at time t' in the decoder and the state at time t on day D in the encoder,/>Time attention coefficient representing the state at time t 'in the decoder and the state at time t' on day d in the encoder, time attention coefficient/>The calculation formula is as follows:
wherein, A correlation coefficient representing a state at time t' of the decoder and a state at time t of the D-th day of the encoder; /(I)A correlation coefficient representing a state at time t 'of the decoder and a state at time t' of the d-th day of the encoder;
all context vectors c t′ corresponding to the time t' in the decoder are calculated by the above formula and are used for calculating the predicted data y t′:
yt′=σ(yt′-1,ct′,st′);
Step 53, using root mean square error RMSE as a loss function of the model, using Adam SGD optimizer to adjust related parameters in the GAT-Seq2Seq model according to the training set, and performing super-parameter optimization through the verification set, thereby obtaining a trained GAT-Seq2Seq model;
And 6, inputting a test set into the trained GAT-Seq2Seq model, predicting the passenger flow of subway stations in each future period, and evaluating the trained GAT-Seq2Seq model according to a prediction result.
2. The short-term prediction method of space-time dual-attention subway passenger flow based on the GAT-Seq2Seq model according to claim 1, wherein in the step 1, the subway station basic information includes: longitude and latitude geographic coordinates of subway stations and subway operation line data; the subway station history passenger flow data comprises: station names or station serial numbers, card swiping inbound time and inbound passenger flow; the historical meteorological data includes: whether to rainfall, whether to fall snow, the rainfall, the snowfall and the temperature; holiday data includes: whether the day of the week and whether the holiday, wherein whether the day of the week is converted into seven parameters using a thermal independent code, namely: whether Monday, tuesday, friday, saturday, and Sunday.
3. The short-term prediction method of the passenger flow of the space-time double-attention subway based on the GAT-Seq2Seq model according to claim 1, wherein the specific process of the step 2 is as follows:
step 21, deleting the data with abnormal or missing subway station names in the related data of the subway passenger flow by adopting an integral removing method, and processing the data with the missing or abnormal subway station names in the subway card swiping and station entering time, station entering passenger flow, rainfall, snowfall and temperature by adopting a forward filling substitution method;
Step 22, sorting historical passenger flow data of subway stations according to time sequence, and setting the time period length l to be 1 hour, dividing one day into 24 time periods, and respectively counting the passenger flow, rainfall and snowfall of the subway stations in each time period;
Step 23, calculating the average temperature in each period according to the acquired temperature data, wherein the maximum temperature and the minimum temperature in each period respectively take the maximum value and the minimum value of all the temperatures in the period;
Step 24, normalizing the subway card-swiping and arrival passenger flow, the rainfall, the snowfall, the average temperature, the highest temperature and the lowest temperature to obtain a data set;
Step 25, the data set is divided into a training set, a verification set and a test set according to the ratio of 3:1:1.
4. The short-term prediction method of the passenger flow of the space-time double-attention subway based on the GAT-Seq2Seq model according to claim 1, wherein the specific process of the step 3 is as follows:
Step 31, taking all subway stations as vertexes of a graph structure, knowing longitude and latitude coordinates of each subway station, and for any vertex A, searching a circular range with the vertex A as a circle center and the radius of 2.5Km, wherein if a certain subway station simultaneously meets the following two conditions: (1) The subway station is in a circular range, (2) a subway operation reachable line exists between the subway station and the vertex A or the actual walking time between the subway station and the vertex A is less than 10 minutes; then consider that there is an edge between the subway station and vertex a, otherwise there is no edge;
step 32, repeating step 31 for all vertices to obtain the subway station map structure G.
5. The short-term prediction method of the passenger flow of the space-time double-attention subway based on the GAT-Seq2Seq model according to claim 1, wherein the specific process of the step 6 is as follows:
step 61, migrating the trained GAT-Seq2Seq model to a test set, determining a prediction step length l to be 1 hour, and predicting subway station passenger flow in a future period by using the trained GAT-Seq2Seq model;
Step 62, calculating a root mean square error RMSE of the predicted value and the actual value for model performance evaluation, wherein the root mean square error RMSE is as follows:
wherein R is the total number of predicted samples, y r is the actual passenger flow volume, To predict passenger flow.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111278744.6A CN113962472B (en) | 2021-10-31 | 2021-10-31 | Short-time prediction method for space-time double-attention subway passenger flow based on GAT-Seq2Seq model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111278744.6A CN113962472B (en) | 2021-10-31 | 2021-10-31 | Short-time prediction method for space-time double-attention subway passenger flow based on GAT-Seq2Seq model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113962472A CN113962472A (en) | 2022-01-21 |
CN113962472B true CN113962472B (en) | 2024-04-19 |
Family
ID=79468546
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111278744.6A Active CN113962472B (en) | 2021-10-31 | 2021-10-31 | Short-time prediction method for space-time double-attention subway passenger flow based on GAT-Seq2Seq model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113962472B (en) |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114004429B (en) * | 2022-01-04 | 2022-04-08 | 苏州元澄科技股份有限公司 | Data processing method and system for constructing digital city |
CN114548777B (en) * | 2022-02-23 | 2024-07-19 | 上海汽车集团股份有限公司 | Shared electric automobile dispatching method and device, electronic equipment and storage medium |
CN114819253A (en) * | 2022-03-02 | 2022-07-29 | 湖北大学 | Urban crowd gathering hotspot area prediction method, system, medium and terminal |
CN117272848B (en) * | 2023-11-22 | 2024-02-02 | 上海随申行智慧交通科技有限公司 | Subway passenger flow prediction method and model training method based on space-time influence |
CN117591919B (en) * | 2024-01-17 | 2024-03-26 | 北京工业大学 | Passenger flow prediction method, passenger flow prediction device, electronic equipment and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020215798A1 (en) * | 2019-04-22 | 2020-10-29 | 中国科学院深圳先进技术研究院 | Method and system for estimating regional passenger traffic in subway station, and electronic device |
CN111860785A (en) * | 2020-07-24 | 2020-10-30 | 中山大学 | Time sequence prediction method and system based on attention mechanism cyclic neural network |
CN112801355A (en) * | 2021-01-20 | 2021-05-14 | 南京航空航天大学 | Data prediction method based on multi-graph fusion space-time attention of long-short-term space-time data |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11158048B2 (en) * | 2019-06-28 | 2021-10-26 | Shandong University Of Science And Technology | CT lymph node detection system based on spatial-temporal recurrent attention mechanism |
-
2021
- 2021-10-31 CN CN202111278744.6A patent/CN113962472B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020215798A1 (en) * | 2019-04-22 | 2020-10-29 | 中国科学院深圳先进技术研究院 | Method and system for estimating regional passenger traffic in subway station, and electronic device |
CN111860785A (en) * | 2020-07-24 | 2020-10-30 | 中山大学 | Time sequence prediction method and system based on attention mechanism cyclic neural network |
CN112801355A (en) * | 2021-01-20 | 2021-05-14 | 南京航空航天大学 | Data prediction method based on multi-graph fusion space-time attention of long-short-term space-time data |
Also Published As
Publication number | Publication date |
---|---|
CN113962472A (en) | 2022-01-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113962472B (en) | Short-time prediction method for space-time double-attention subway passenger flow based on GAT-Seq2Seq model | |
CN110570651B (en) | Road network traffic situation prediction method and system based on deep learning | |
Liu et al. | Contextualized spatial–temporal network for taxi origin-destination demand prediction | |
CN109658695B (en) | Multi-factor short-term traffic flow prediction method | |
CN108564790B (en) | Urban short-term traffic flow prediction method based on traffic flow space-time similarity | |
CN109448361B (en) | Resident traffic travel flow prediction system and prediction method thereof | |
CN110956807B (en) | Highway flow prediction method based on combination of multi-source data and sliding window | |
CN112489426B (en) | Urban traffic flow space-time prediction scheme based on graph convolution neural network | |
CN110555544B (en) | Traffic demand estimation method based on GPS navigation data | |
CN110163449B (en) | Motor vehicle pollution discharge monitoring node deployment method based on active space-time diagram convolution | |
CN112863182B (en) | Cross-modal data prediction method based on transfer learning | |
CN115440032A (en) | Long-term and short-term public traffic flow prediction method | |
CN111242395B (en) | Method and device for constructing prediction model for OD (origin-destination) data | |
CN115545758A (en) | Method and system for self-adaptive incremental site selection of urban service facilities | |
CN115204477A (en) | Bicycle flow prediction method of context awareness graph recursive network | |
CN117494034A (en) | Air quality prediction method based on traffic congestion index and multi-source data fusion | |
Zhao et al. | Celltrademap: Delineating trade areas for urban commercial districts with cellular networks | |
CN115412857A (en) | Resident travel information prediction method | |
CN116796904A (en) | Method, system, electronic equipment and medium for predicting new line passenger flow of rail transit | |
CN113537569B (en) | Short-term bus passenger flow prediction method and system based on weight stacking decision tree | |
CN116993391A (en) | Site type shared bicycle system use demand prediction method | |
CN117037461A (en) | Road network traffic jam prediction method based on multi-weight graph three-dimensional convolution | |
CN114330871A (en) | Method for predicting urban road conditions by combining public transport operation data with GPS data | |
AT&T | ||
CN114139773A (en) | Public traffic flow prediction method based on space-time graph convolutional network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |